Human engagement state recognition for autonomous functioning of a robot in human-robot conversation
MetadataShow full item record
The goal of this thesis was to develop a model to classify the different states of engagement. We took on the definition of engagement as the process by which interactors start, maintain and end their perceived connection to each other during interaction and included the state where an interactor does not have or no longer has the intention to interact. Based on this, four states could be distinguished: no interest, intention to interact (or interest), engaged and ending interaction. The purpose of developing this model was to contribute to improving an informative conversation between human and robot by improving the way a robot determines who to engage with or pay attention to. Since engagement behaviour is not well understood in the human-human context, despite its apparent significance, we looked further into the research done both in human-human and human-robot interaction. Based on this, we have composed a set of features and set up a Naive Bayes classifier to classify the states of engagement. The features used are: distance from the robot, facing direction, gaze, position, sound direction and velocity. The model can classify one person at a time, however the system is designed with the possibility to expand it for multiple people as well as additional features. We intend to both choose the features and design the model in a way that it can be used regardless of the robot or platform as much as possible. However, to allow testing and to have a system that can be used in practice, we take the humanoid robot Pepper, developed by Softbank, as our main platform. This has given some limitations as to what features are chosen and to how the model is implemented. Evaluation of the model gives promising results for the overall model and the states no interest, intention to interact and engaged, however the model performs badly for the state ending interaction. We discuss for the latter state specifically and for the model in general possibilities for improvement.